Developer Experience in the Age of AI Coding Agents (Agents column)
I saw something interesting. speech Max Kanat-Alexander, Executive Distinguished Engineer at Capital One, where he talked about DevEx in the era of AI agents or how not to drown in the already existing legacy and newly created legacy generated by agents in accelerated mode:) Max previously worked at Google Tech Lead. (Code Health) Principal Staff Engineer on LinkedIn (Developer Productivity). He also wrote books.Code SimplicityandUnderstanding Software" which I have not yet read:)
The main points of the report are such
1️ Don't fight with Training Set
Use standard tools. If you write your package manager or use rare language, the agent will be dumb. It is trained in open source standards. The more “boring” and standard your stack, the smarter the AI runs on it.
2️⃣ CLI > GUI
Agents need an API and a CLI, not a browser. Getting an agent to click through a GUI is expensive and unreliable. If the tool has a text interface, the agent will handle it faster and more accurately.
3Tests should be deterministic
The agent doesn't say anything.*500 Internal Error*\. He needs clear messages from validators. Investing in quality error messages in tests and linters is an investment in an agent’s autonomy. Critical that tests should run fast (30 seconds, not 20 minute). The agent runs them in a loop. Slow CI will kill the agent's productivity.
4Document "why", not "what"
The agent sees the code and understands what He does. But he hasn't been to your rallies and can't read minds.
Documentation now needs to be written contextBusiness objectives, external constraints, form of data at the input. What's not in the code.
5The Problem of Code Review (A vicious cycle)
Writing code becomes reading. The number of PRs is growing exponentially.
But if you have a weak review process, you'll start stamping. (LGTM) Bad code from agents. The code base degrades, it becomes more difficult for the agent to work in it, he writes even more nonsense → it turns out a vicious circle. The solution is to distribute the burden on the audit (Don't send it all to the "somebody look" channel.) And hold the bar tight.
What does that mean for development? - Legasi refactoring is mandatory.. If a person can’t understand the structure of a project without “secret knowledge,” the agent is hallucinating. A good code structure is now an economic necessity. - A paradigm shift.. We are moving from code writing to verification. The ability to quickly read and validate someone else’s code becomes more important than the ability to type quickly. - The golden rule.. All that is good for an AI agent (Quick tests, clear errors, standard tools)It's good for the human. Even if AI disappears tomorrow, that investment will pay off for people. It's funny that it looks a lot likemoralismIt is a universal ethical principle that says, “Do unto others as you would be done unto you.”
#Engineering #AI #Metrics #Software #DevEx #Productivity #DevOps #Architecture #Culture #Engineering #ML #SystemDesign